TASK 35 — Leaving the room the robot started in

TASK 33 gave the executor an ordered plan and TASK 34 taught it to drive a passage. Both were measured on scenes where every destination was visible from somewhere in one room. home_building_1 q5 is not that scene: its first destination is a nightstand at (-1.90, -12.92), thirteen metres from the start and through two doorways, and every attempt at it failed. This task is about why, and it turns out the model was never the problem.

What the failing runs actually showed

Leg 1 of home_building_1 q5, twice, before any change here:

	hm1_q2_reserve (1st)	hm1_q2_reserve (2nd)
ground covered	x∈[-5.1, 0], y∈[-2.0, 1.3]	x∈[0, 7.3], y∈[-2.9, -7.6]
ended because	first return to a visited pose	9 steps used
time used / allotted	183 / 400 s	212 / 400 s

Both legs died with more than three minutes of their own slice unspent, and both died of a proxy for time rather than time itself.

The model's reasoning was sound throughout. Its explore.why named a visible doorway on seven of nine calls — "the wide framed doorway on the left of image 0 opens into a bedroom-like room with a bed and lamp" — and the leg walked 20 m for 7 m of net displacement without entering a single room.

Four causes, each measured

1. Circling detection killed the search. revisited() was written for a leg that had already bound its target and was walking the ring around it, where coming back means there is nothing more to gain. In a multi-room search it means the opposite: backing out of a dead end and returning to the hall to try another door is what searching a building is. It now counts returns instead of ending on the first one, up to MAX_LOOPS = 3, and tells the model it has looped.

2. Nothing remembered which door had been tried. The model re-reasons from scratch on every call and kept nominating the door it had just come out of; the terrain cannot tell two doors apart. spent records each departure as (from, unit direction) and already_tried() matches within REVISIT_M and SPENT_CONE_DEG = 40°. A repeat is scored down by SPENT_PENALTY = 0.25, not forbidden — a room with one exit must keep that exit reachable.

3. reach · cos(Δ) was structurally unable to leave a room. This is the important one. A doorway is a short-reach bearing: an 0.9 m opening onto a small room stops the ray a couple of metres in, while the corridor beside it runs five or six. Maximising reach therefore prefers the open space to the door every time, by construction. In the nine-step leg above, five steps were swung 15–30° off what the model asked and all five landed on bearings reaching 4.4–6.8 m. Reach is now a gate, not an objective: among bearings clearing MIN_EXPLORE_M, the one nearest what was asked wins. The old rule survives as a fallback for a vehicle hemmed in on all sides, so loft's "heading 270° reaches 0.00 m" fix is not lost.

4. A bearing cannot express "through that door". Prompt v6-way-out adds a way field — the opening, boxed like any other object — and lift_way() places it with the same _lift_xy that already places targets, anchors and gates. Built by surgery on v5, so v5 stays byte-for-byte exact; the 117 cached replies and every offline script are keyed to it.

--goto-steps went 9 → 20, since the leg's share of --budget is the real governor and the step cap is a safety valve.

The cap that v6 needed immediately

The first v6 run used the way-out branch on 7 of 8 exploration steps — the mechanism works — and drove the robot into a bedroom, where it inspected both nightstands, found no clock, and correctly left. No previous run had entered any room at all.

But four of seven lifts came back beyond 9 m, and the first, at 13.69 m, landed at (+8.57, -10.68) — outside the entire extent the reference trajectory ever visits, which is what aimed the leg east from its first move. The cause is the thing that makes the field work close up: the scanner sees through an opening and returns whatever stands behind it. Near to hand that is usefully past the threshold; far away it is the room after next.

WAY_MAX_M = 7.0 keeps the bearing and drops the range. Replaying the run's own eleven lifts, seven are clamped and the one that drove the robot into the bedroom (4.47 m) is untouched.

Where this leaves the scenes

home_building_2 q1, driven end to end against trajectory_q4.ply:

	GT	ours	gap
destination 1 (magazine on ottoman)	reversal at (+0.83, +3.58)	bound (+0.58, +4.16)	0.52 m
destination 2 (plant on dressing table)	end (-4.07, +11.66)	bound (-3.37, +10.57)	1.30 m

The robot stopped 0.95 m from its binding where the converter's best legal point in the whole terrain frame was 0.92 m — three centimetres off the platform's floor, which obstacleDisThre and the vehicle footprint set. The reference trajectory's own endpoint is 1.30 m from that binding, so the run parks closer to the object than GT does. 2/2.

home_building_1 q5 still fails leg 1, and the last run failed it in a new way worth recording: the model matched the feature and dropped the head noun. Asked for "the nightstand with a clock on it", it answered NOT_VISIBLE three times in a row while looking straight at two lamp-topped nightstands — exactly what branch (b) asks for — then found a clock on a TV sideboard and reported visible: true at target_type: nightstand, binding 12.10 m from the right answer. Its own evidence says "sideboard", "under the wall-mounted TV", "bedside-type", and "the only clock" — a scarcity argument overriding the noun. Branch (b) covers "right type, feature missing → NOT_VISIBLE" and says nothing about the converse. Not fixed here.

Files

• scripts/approach_loop.py — lift_way, already_tried, rewritten explore_direction, loop counting; WAY_MAX_M, SPENT_CONE_DEG, SPENT_PENALTY, MAX_LOOPS
• scripts/vlm_probe.py — _make_v6, _WAY_BRANCH, _WAY_FIELDS
• scripts/execute_plan.py — --goto-steps default
• scripts/show_run.py — new; reads a run directory back as one screen per step
• tests/test_execute_plan.py — +52 tests, on the recorded numbers above

Still open

• No responder. Nothing here is submittable; a submission is a ROS node under ai_module/ reacting to /challenge_question.
• The feature-over-noun false positive above.
• A destination thirteen metres away through two doorways is still unreached.
• KEEPOUT_M = 1.2 refuted in TASK 31, still a disc rather than a corridor.